K-Bench
Models respond across multi-turn mental-health risk conversations involving suicide, self-harm, domestic violence, substance misuse, and no-risk cases.
The public leaderboard compares 125 model configurations from 33 base models and 14 providers on 200 vignettes. Six clinicians rated 151 transcripts for judge calibration.
- Made by
- Kivira Health and University of Roehampton
- Released
- September 2026
- Kind
- Wellbeing and sycophancy
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Open-ended
- Tasks
- Public, plus a private test set
- Measures
- Propensity
- Behavior
- Misaligned with values
- Subject
- Clinical medicine, Psychology
- Language
- English
- Interaction
- Multi-turn
- Items from
- Generated
- Refresh
- One fixed set
- Run by
- A host
- In views
- Biology and medicine, Content safety
- Strongest model tested
- Claude Fable 5 #12 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 20th, 2026
GPT-5.5 leads with 99% among models released by July 2026 GPT-5.5 leads with 99% among models released by July 2026
Best score per entry from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 13 more
- 99.0 GPT-5.5
- 99.0 GPT-5.2
- 98.6 Kimi K2.5
- 98.6 Claude Opus 4.8
- 98.6 Claude Fable 5
- 98.5 GPT-5.4
- 98.5 Claude Haiku 4.5
- 98.4 Kimi K2.6
- 98.4 stealth/ox-alpha
- 98.3 Grok 4.20
- 98.1
Gemini 3.1 Flash-Lite
- 98.0 GPT-5.6 Terra
- 98.0 GPT-5.6 Sol
- 97.8 nvidia/nemotron-3-ultra-550b-a55b
- 97.7
Gemini 3.5 Flash
- 97.7
Qwen3-32B
- 97.6
Gemini 2.5 Flash
- 97.6
DeepSeek-V4-Flash
- 97.6 GPT-5.6 Luna
- 97.5 GLM-5.2
- 97.5
Gemini 3.1 Pro
- 97.2
Gemini 2.5 Flash-Lite
- 97.0
Mistral Medium 3.5
- 97.0 nvidia/nemotron-3-super-120b-a12b
- 96.5 Grok 4.3
- 96.4 mistralai/mistral-small-2603
- 96.4
Llama 4 Maverick
- 96.0 Granite 4.1 8B
- 95.7
Qwen3-8B
- 95.7 GPT-4o mini
- 95.6
Phi-4
- 94.5 Grok 4.5
- 89.8 ibm-granite/granite-4.0-h-micro
- 1 GPT-5.5 Default system prompt v1, low 99%
- 2 GPT-5.2 Therapeutic prompt v0, none 99%
- 3 Kimi K2.5 Default system prompt v1, default 98.6%
- 4 Claude Opus 4.8 Default system prompt v1, low 98.6%
- 5 Claude Fable 5 Default system prompt v1, high 98.6%
- 6 GPT-5.4 Therapeutic prompt v0, low 98.5%
- 7 Claude Haiku 4.5 Therapeutic prompt v0, none 98.5%
- 8 Kimi K2.6 Default system prompt v1, default 98.4%
- 9 stealth/ox-alpha Default system prompt v1, low 98.4%
- 10 Grok 4.20 Default system prompt v1, none 98.3%
- 11
Gemini 3.1 Flash-Lite Default system prompt v1, default_minimal 98.1%
- 12 GPT-5.6 Terra Therapeutic prompt v0, none 98%
- 13 GPT-5.6 Sol Default system prompt v1, none 98%
- 14 nvidia/nemotron-3-ultra-550b-a55b Therapeutic prompt v0, default_high 97.8%
- 15
Gemini 3.5 Flash Therapeutic prompt v0, default_medium 97.7%
- 16
Qwen3-32B Therapeutic prompt v0, high 97.7%
- 17
Gemini 2.5 Flash Sep 2025, Default system prompt v1, none 97.6%
- 18
DeepSeek-V4-Flash Therapeutic prompt v0, high 97.6%
- 19 GPT-5.6 Luna Default system prompt v1, none 97.6%
- 20 GLM-5.2 Therapeutic prompt v0, high 97.5%
- 21
Gemini 3.1 Pro Therapeutic prompt v0, default_medium 97.5%
- 22
Gemini 2.5 Flash-Lite Sep 2025, Default system prompt v1, none 97.2%
- 23
Mistral Medium 3.5 Therapeutic prompt v0, none 97%
- 24 nvidia/nemotron-3-super-120b-a12b Default system prompt v1, default_medium 97%
- 25 Grok 4.3 Therapeutic prompt v0, high 96.5%
- 26 mistralai/mistral-small-2603 Therapeutic prompt v0, none 96.4%
- 27
Llama 4 Maverick Therapeutic prompt v0, none 96.4%
- 28 Granite 4.1 8B Therapeutic prompt v0, none 96%
- 29
Qwen3-8B Default system prompt v1, default 95.7%
- 30 GPT-4o mini Default system prompt v1, none 95.7%
- 31
Phi-4 Therapeutic prompt v0, none 95.6%
- 32 Grok 4.5 Default system prompt v1, high 94.5%
- 33 ibm-granite/granite-4.0-h-micro Therapeutic prompt v0, none 89.8%
Source: Official K-Bench leaderboard
- 99.0 GPT-5.5
- 99.0 GPT-5.2
- 98.6 Kimi K2.5
- 98.6 Claude Opus 4.8
- 98.6 Claude Fable 5
- 98.5 GPT-5.4
- 98.5 Claude Haiku 4.5
- 98.4 Kimi K2.6
- 98.3 Grok 4.20
- 98.1
Gemini 3.1 Flash-Lite
- 98.0 GPT-5.6 Terra
- 98.0 GPT-5.6 Sol
- 97.7
Gemini 3.5 Flash
- 97.7
Qwen3-32B
- 97.6
Gemini 2.5 Flash
- 97.6
DeepSeek-V4-Flash
- 97.6 GPT-5.6 Luna
- 97.5 GLM-5.2
- 97.5
Gemini 3.1 Pro
- 97.2
Gemini 2.5 Flash-Lite
- 97.0
Mistral Medium 3.5
- 96.5 Grok 4.3
- 96.4
Llama 4 Maverick
- 96.0 Granite 4.1 8B
- 95.7
Qwen3-8B
- 95.7 GPT-4o mini
- 95.6
Phi-4
- 94.5 Grok 4.5
- 1 GPT-5.5 Default system prompt v1, low 99%
- 2 GPT-5.2 Therapeutic prompt v0, none 99%
- 3 Kimi K2.5 Default system prompt v1, default 98.6%
- 4 Claude Opus 4.8 Default system prompt v1, low 98.6%
- 5 Claude Fable 5 Default system prompt v1, high 98.6%
- 6 GPT-5.4 Therapeutic prompt v0, low 98.5%
- 7 Claude Haiku 4.5 Therapeutic prompt v0, none 98.5%
- 8 Kimi K2.6 Default system prompt v1, default 98.4%
- 9 Grok 4.20 Default system prompt v1, none 98.3%
- 10
Gemini 3.1 Flash-Lite Default system prompt v1, default_minimal 98.1%
- 11 GPT-5.6 Terra Therapeutic prompt v0, none 98%
- 12 GPT-5.6 Sol Default system prompt v1, none 98%
- 13
Gemini 3.5 Flash Therapeutic prompt v0, default_medium 97.7%
- 14
Qwen3-32B Therapeutic prompt v0, high 97.7%
- 15
Gemini 2.5 Flash Sep 2025, Default system prompt v1, none 97.6%
- 16
DeepSeek-V4-Flash Therapeutic prompt v0, high 97.6%
- 17 GPT-5.6 Luna Default system prompt v1, none 97.6%
- 18 GLM-5.2 Therapeutic prompt v0, high 97.5%
- 19
Gemini 3.1 Pro Therapeutic prompt v0, default_medium 97.5%
- 20
Gemini 2.5 Flash-Lite Sep 2025, Default system prompt v1, none 97.2%
- 21
Mistral Medium 3.5 Therapeutic prompt v0, none 97%
- 22 Grok 4.3 Therapeutic prompt v0, high 96.5%
- 23
Llama 4 Maverick Therapeutic prompt v0, none 96.4%
- 24 Granite 4.1 8B Therapeutic prompt v0, none 96%
- 25
Qwen3-8B Default system prompt v1, default 95.7%
- 26 GPT-4o mini Default system prompt v1, none 95.7%
- 27
Phi-4 Therapeutic prompt v0, none 95.6%
- 28 Grok 4.5 Default system prompt v1, high 94.5%
Source: Official K-Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 99% | Default system prompt v1, low | Apr 2026 | Official leaderboard |
| 2 | GPT-5.2 | 99% | Therapeutic prompt v0, none | Dec 2025 | Official leaderboard |
| 3 | Kimi K2.5 | 98.6% | Default system prompt v1, default | Jan 2026 | Official leaderboard |
| 4 | Claude Opus 4.8 | 98.6% | Default system prompt v1, low | May 2026 | Official leaderboard |
| 5 | Claude Fable 5 | 98.6% | Default system prompt v1, high | Jun 2026 | Official leaderboard |
| 6 | GPT-5.4 | 98.5% | Therapeutic prompt v0, low | Mar 2026 | Official leaderboard |
| 7 | Claude Haiku 4.5 | 98.5% | Therapeutic prompt v0, none | Oct 2025 | Official leaderboard |
| 8 | Kimi K2.6 | 98.4% | Default system prompt v1, default | Apr 2026 | Official leaderboard |
| 9 | stealth/ox-alpha | 98.4% | Default system prompt v1, low | Official leaderboard | |
| 10 | Grok 4.20 | 98.3% | Default system prompt v1, none | Feb 2026 | Official leaderboard |
| 11 | | 98.1% | Default system prompt v1, default_minimal | Mar 2026 | Official leaderboard |
| 12 | GPT-5.6 Terra | 98% | Therapeutic prompt v0, none | Jul 2026 | Official leaderboard |
| 13 | GPT-5.6 Sol | 98% | Default system prompt v1, none | Jul 2026 | Official leaderboard |
| 14 | nvidia/nemotron-3-ultra-550b-a55b | 97.8% | Therapeutic prompt v0, default_high | Official leaderboard | |
| 15 | | 97.7% | Therapeutic prompt v0, default_medium | May 2026 | Official leaderboard |
| 16 | | 97.7% | Therapeutic prompt v0, high | Apr 2025 | Official leaderboard |
| 17 | | 97.6% | Default system prompt v1, none | Sep 2025 | Official leaderboard |
| 18 | | 97.6% | Therapeutic prompt v0, high | Apr 2026 | Official leaderboard |
| 19 | GPT-5.6 Luna | 97.6% | Default system prompt v1, none | Jul 2026 | Official leaderboard |
| 20 | GLM-5.2 | 97.5% | Therapeutic prompt v0, high | Jun 2026 | Official leaderboard |
| 21 | | 97.5% | Therapeutic prompt v0, default_medium | Feb 2026 | Official leaderboard |
| 22 | | 97.2% | Default system prompt v1, none | Sep 2025 | Official leaderboard |
| 23 | | 97% | Therapeutic prompt v0, none | Apr 2026 | Official leaderboard |
| 24 | nvidia/nemotron-3-super-120b-a12b | 97% | Default system prompt v1, default_medium | Official leaderboard | |
| 25 | Grok 4.3 | 96.5% | Therapeutic prompt v0, high | Apr 2026 | Official leaderboard |
| 26 | mistralai/mistral-small-2603 | 96.4% | Therapeutic prompt v0, none | Official leaderboard | |
| 27 | | 96.4% | Therapeutic prompt v0, none | Apr 2025 | Official leaderboard |
| 28 | Granite 4.1 8B | 96% | Therapeutic prompt v0, none | Apr 2026 | Official leaderboard |
| 29 | | 95.7% | Default system prompt v1, default | Apr 2025 | Official leaderboard |
| 30 | GPT-4o mini | 95.7% | Default system prompt v1, none | Jul 2024 | Official leaderboard |
| 31 | | 95.6% | Therapeutic prompt v0, none | Dec 2024 | Official leaderboard |
| 32 | Grok 4.5 | 94.5% | Default system prompt v1, high | Jul 2026 | Official leaderboard |
| 33 | ibm-granite/granite-4.0-h-micro | 89.8% | Therapeutic prompt v0, none | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5.5 | 99% | Default system prompt v1, low | Apr 2026 | Official leaderboard |
| 2 | GPT-5.2 | 99% | Therapeutic prompt v0, none | Dec 2025 | Official leaderboard |
| 3 | Kimi K2.5 | 98.6% | Default system prompt v1, default | Jan 2026 | Official leaderboard |
| 4 | Claude Opus 4.8 | 98.6% | Default system prompt v1, low | May 2026 | Official leaderboard |
| 5 | Claude Fable 5 | 98.6% | Default system prompt v1, high | Jun 2026 | Official leaderboard |
| 6 | GPT-5.4 | 98.5% | Therapeutic prompt v0, low | Mar 2026 | Official leaderboard |
| 7 | Claude Haiku 4.5 | 98.5% | Therapeutic prompt v0, none | Oct 2025 | Official leaderboard |
| 8 | Kimi K2.6 | 98.4% | Default system prompt v1, default | Apr 2026 | Official leaderboard |
| 9 | Grok 4.20 | 98.3% | Default system prompt v1, none | Feb 2026 | Official leaderboard |
| 10 | | 98.1% | Default system prompt v1, default_minimal | Mar 2026 | Official leaderboard |
| 11 | GPT-5.6 Terra | 98% | Therapeutic prompt v0, none | Jul 2026 | Official leaderboard |
| 12 | GPT-5.6 Sol | 98% | Default system prompt v1, none | Jul 2026 | Official leaderboard |
| 13 | | 97.7% | Therapeutic prompt v0, default_medium | May 2026 | Official leaderboard |
| 14 | | 97.7% | Therapeutic prompt v0, high | Apr 2025 | Official leaderboard |
| 15 | | 97.6% | Default system prompt v1, none | Sep 2025 | Official leaderboard |
| 16 | | 97.6% | Therapeutic prompt v0, high | Apr 2026 | Official leaderboard |
| 17 | GPT-5.6 Luna | 97.6% | Default system prompt v1, none | Jul 2026 | Official leaderboard |
| 18 | GLM-5.2 | 97.5% | Therapeutic prompt v0, high | Jun 2026 | Official leaderboard |
| 19 | | 97.5% | Therapeutic prompt v0, default_medium | Feb 2026 | Official leaderboard |
| 20 | | 97.2% | Default system prompt v1, none | Sep 2025 | Official leaderboard |
| 21 | | 97% | Therapeutic prompt v0, none | Apr 2026 | Official leaderboard |
| 22 | Grok 4.3 | 96.5% | Therapeutic prompt v0, high | Apr 2026 | Official leaderboard |
| 23 | | 96.4% | Therapeutic prompt v0, none | Apr 2025 | Official leaderboard |
| 24 | Granite 4.1 8B | 96% | Therapeutic prompt v0, none | Apr 2026 | Official leaderboard |
| 25 | | 95.7% | Default system prompt v1, default | Apr 2025 | Official leaderboard |
| 26 | GPT-4o mini | 95.7% | Default system prompt v1, none | Jul 2024 | Official leaderboard |
| 27 | | 95.6% | Therapeutic prompt v0, none | Dec 2024 | Official leaderboard |
| 28 | Grok 4.5 | 94.5% | Default system prompt v1, high | Jul 2026 | Official leaderboard |
Rankings by part
We read 8 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Combined risk, Clinical judgment and Risk exploration. GPT-4o mini leads 4 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Combined risk
28 models we track have a score here.
- 1 Claude Fable 5 ↑4 96.1%
- 2 GPT-5.5 ↓1 96%
- 3 GPT-5.2 ↓1 95.2%
- 4 GPT-5.6 Sol ↑8 95%
- 5 GPT-5.6 Terra ↑6 94.6%
- 6 GPT-5.4 94.4%
- 7 Claude Opus 4.8 ↓3 94.3%
- 8 GPT-5.6 Luna ↑9 94%
- 9 Kimi K2.5 ↓6 94%
- 10 Claude Haiku 4.5 ↓3 93.4%
Clinical judgment
28 models we track have a score here.
- 1 GPT-4o mini ↑25 100%
- 2
Phi-4 ↑25 100%
- 3
Llama 4 Maverick ↑20 100%
- 4
Qwen3-32B ↑10 100%
- 5
Qwen3-8B ↑20 100%
- 6
Gemini 2.5 Flash-Lite Sep 2025 ↑14 100%
- 7
Gemini 2.5 Flash Sep 2025 ↑8 100%
- 8 Claude Haiku 4.5 ↓1 100%
- 9 GPT-5.2 ↓7 100%
- 10 Kimi K2.5 ↓7 100%
Risk exploration
28 models we track have a score here.
- 1 Claude Fable 5 ↑4 93.1%
- 2 GPT-5.5 ↓1 91.7%
- 3 GPT-5.2 ↓1 90.8%
- 4 GPT-5.6 Sol ↑8 89.5%
- 5 GPT-5.4 ↑1 89.4%
- 6 Claude Opus 4.8 ↓2 89.4%
- 7 GPT-5.6 Terra ↑4 89.3%
- 8
Gemini 3.1 Flash-Lite ↑2 88.1%
- 9 Kimi K2.5 ↓6 88%
- 10 GPT-5.6 Luna ↑7 88%
Ethical reasoning
28 models we track have a score here.
- 1 Claude Haiku 4.5 ↑6 100%
- 2 GPT-5.4 ↑4 100%
- 3 GPT-5.5 ↓2 100%
- 4 GPT-5.6 Luna ↑13 100%
- 5 GPT-5.6 Sol ↑7 100%
- 6 GPT-5.6 Terra ↑5 100%
- 7 Claude Fable 5 ↓2 99.9%
- 8 Kimi K2.5 ↓5 99.8%
- 9 Kimi K2.6 ↓1 99.8%
- 10 Grok 4.3 ↑12 99.8%
Supportive conversation
28 models we track have a score here.
- 1 GPT-4o mini ↑25 100%
- 2
Phi-4 ↑25 100%
- 3
Llama 4 Maverick ↑20 100%
- 4
Qwen3-32B ↑10 100%
- 5
Gemini 2.5 Flash-Lite Sep 2025 ↑15 100%
- 6
Gemini 2.5 Flash Sep 2025 ↑9 100%
- 7 Claude Haiku 4.5 100%
- 8 GPT-5.2 ↓6 100%
- 9 Grok 4.20 100%
- 10
Gemini 3.1 Pro ↑9 100%
Psychological knowledge
28 models we track have a score here.
- 1 GPT-4o mini ↑25 100%
- 2
Phi-4 ↑25 100%
- 3
Qwen3-32B ↑11 100%
- 4
Qwen3-8B ↑21 100%
- 5
Gemini 2.5 Flash-Lite Sep 2025 ↑15 100%
- 6 Claude Haiku 4.5 ↑1 100%
- 7 GPT-5.2 ↓5 100%
- 8 Kimi K2.5 ↓5 100%
- 9
Gemini 3.1 Flash-Lite ↑1 100%
- 10 GPT-5.4 ↓4 100%
Cultural competence
28 models we track have a score here.
- 1
Qwen3-32B ↑13 100%
- 2
Gemini 2.5 Flash Sep 2025 ↑13 100%
- 3 Claude Haiku 4.5 ↑4 100%
- 4 GPT-5.2 ↓2 100%
- 5 Grok 4.20 ↑4 100%
- 6 GPT-5.5 ↓5 100%
- 7
DeepSeek-V4-Flash ↑9 100%
- 8 Granite 4.1 8B ↑16 100%
- 9
Gemini 3.5 Flash ↑4 100%
- 10 Claude Opus 4.8 ↓6 100%
Autonomy and boundaries
28 models we track have a score here.
- 1 GPT-4o mini ↑25 100%
- 2
Phi-4 ↑25 100%
- 3
Llama 4 Maverick ↑20 100%
- 4
Qwen3-32B ↑10 100%
- 5
Qwen3-8B ↑20 100%
- 6
Gemini 2.5 Flash-Lite Sep 2025 ↑14 100%
- 7
Gemini 2.5 Flash Sep 2025 ↑8 100%
- 8 Claude Haiku 4.5 ↓1 100%
- 9 GPT-5.2 ↓7 100%
- 10 Kimi K2.5 ↓7 100%
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official K-Bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.5 | Default system prompt v1, low | 99% | |||
| GPT-5.2 | Therapeutic prompt v0, none | 99% | |||
| GPT-5.5 | Default system prompt v1, default_medium | 98.9% | |||
| GPT-5.5 | Therapeutic prompt v0, low | 98.7% | |||
| GPT-5.5 | Therapeutic prompt v0, default_medium | 98.7% | |||
| GPT-5.2 | Therapeutic prompt v0, high | 98.6% | |||
| Kimi K2.5 | Default system prompt v1, default | 98.6% | |||
| GPT-5.2 | Default system prompt v1, none | 98.6% | |||
| Claude Opus 4.8 | Default system prompt v1, low | 98.6% | |||
| Claude Fable 5 | Default system prompt v1, high | 98.6% | |||
| GPT-5.2 | Default system prompt v1, high | 98.5% | |||
| GPT-5.4 | Therapeutic prompt v0, low | 98.5% | |||
| Claude Fable 5 | Default system prompt v1, low | 98.5% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, none | 98.5% | |||
| Kimi K2.5 | Therapeutic prompt v0, default | 98.5% | |||
| Kimi K2.6 | Default system prompt v1, default | 98.4% | |||
| GPT-5.4 | Default system prompt v1, none | 98.4% | |||
| GPT-5.5 | Therapeutic prompt v0, none | 98.4% | |||
| GPT-5.4 | Default system prompt v1, low | 98.4% | |||
| GPT-5.4 | Therapeutic prompt v0, high | 98.4% | |||
| GPT-5.4 | Therapeutic prompt v0, none | 98.4% | |||
| GPT-5.5 | Therapeutic prompt v0, high | 98.4% | |||
| Claude Haiku 4.5 | Default system prompt v1, high | 98.3% | |||
| Claude Opus 4.8 | Default system prompt v1, high | 98.3% | |||
| Grok 4.20 | Default system prompt v1, none | 98.3% | |||
| GPT-5.5 | Default system prompt v1, high | 98.3% | |||
| Claude Haiku 4.5 | Default system prompt v1, low | 98.2% | |||
| Kimi K2.6 | Default system prompt v1, high | 98.2% | |||
| GPT-5.4 | Default system prompt v1, high | 98.2% | |||
| Default system prompt v1, default_minimal | 98.1% | ||||
| Kimi K2.6 | Therapeutic prompt v0, default | 98.1% | |||
| Claude Opus 4.8 | Default system prompt v1, none | 98.1% | |||
| Claude Fable 5 | Therapeutic prompt v0, low | 98.1% | |||
| GPT-5.5 | Default system prompt v1, none | 98% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, none | 98% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, low | 98% | |||
| GPT-5.6 Sol | Default system prompt v1, none | 98% | |||
| GPT-5.6 Terra | Default system prompt v1, default_medium | 98% | |||
| GPT-5.6 Terra | Default system prompt v1, none | 97.9% | |||
| Claude Opus 4.8 | Therapeutic prompt v0, low | 97.9% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, default_medium | 97.8% | |||
| Claude Opus 4.8 | Therapeutic prompt v0, none | 97.8% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, high | 97.8% | |||
| Claude Fable 5 | Therapeutic prompt v0, high | 97.8% | |||
| Therapeutic prompt v0, default_medium | 97.7% | ||||
| Claude Haiku 4.5 | Default system prompt v1, none | 97.7% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, high | 97.7% | |||
| Therapeutic prompt v0, high | 97.7% | ||||
| Default system prompt v1, none | 97.6% | ||||
| GPT-5.6 Sol | Therapeutic prompt v0, none | 97.6% | |||
| Therapeutic prompt v0, high | 97.6% | ||||
| GPT-5.6 Sol | Default system prompt v1, default_medium | 97.6% | |||
| GPT-5.6 Luna | Default system prompt v1, none | 97.6% | |||
| Therapeutic prompt v0, high | 97.5% | ||||
| Claude Opus 4.8 | Therapeutic prompt v0, high | 97.5% | |||
| GPT-5.6 Sol | Default system prompt v1, high | 97.5% | |||
| GLM-5.2 | Therapeutic prompt v0, high | 97.5% | |||
| Default system prompt v1, high | 97.5% | ||||
| Therapeutic prompt v0, default | 97.5% | ||||
| Therapeutic prompt v0, default_medium | 97.5% | ||||
| GPT-5.6 Terra | Default system prompt v1, high | 97.5% | |||
| GPT-5.6 Luna | Therapeutic prompt v0, none | 97.4% | |||
| GLM-5.2 | Default system prompt v1, default_high | 97.4% | |||
| Default system prompt v1, low | 97.4% | ||||
| Default system prompt v1, high | 97.3% | ||||
| Grok 4.20 | Therapeutic prompt v0, none | 97.3% | |||
| Kimi K2.6 | Therapeutic prompt v0, high | 97.3% | |||
| Grok 4.20 | Therapeutic prompt v0, high | 97.2% | |||
| Default system prompt v1, default_medium | 97.2% | ||||
| Default system prompt v1, none | 97.2% | ||||
| Therapeutic prompt v0, default_minimal | 97.2% | ||||
| Default system prompt v1, high | 97.2% | ||||
| Therapeutic prompt v0, none | 97.1% | ||||
| GPT-5.6 Luna | Default system prompt v1, default_medium | 97.1% | |||
| Default system prompt v1, default | 97% | ||||
| Therapeutic prompt v0, none | 97% | ||||
| GPT-5.6 Luna | Default system prompt v1, high | 97% | |||
| GPT-5.6 Sol | Therapeutic prompt v0, default_medium | 97% | |||
| Therapeutic prompt v0, none | 97% | ||||
| GPT-5.6 Sol | Therapeutic prompt v0, high | 97% | |||
| Default system prompt v1, default_medium | 97% | ||||
| Default system prompt v1, none | 97% | ||||
| Grok 4.20 | Default system prompt v1, high | 96.8% | |||
| Default system prompt v1, low | 96.8% | ||||
| Therapeutic prompt v0, none | 96.8% | ||||
| GPT-5.6 Luna | Therapeutic prompt v0, default_medium | 96.7% | |||
| GPT-5.6 Luna | Therapeutic prompt v0, high | 96.6% | |||
| Grok 4.3 | Therapeutic prompt v0, high | 96.5% | |||
| Therapeutic prompt v0, none | 96.4% | ||||
| Default system prompt v1, high | 96.3% | ||||
| Granite 4.1 8B | Therapeutic prompt v0, none | 96% | |||
| Grok 4.3 | Therapeutic prompt v0, default_low | 96% | |||
| Default system prompt v1, none | 95.9% | ||||
| Default system prompt v1, default | 95.7% | ||||
| GPT-4o mini | Default system prompt v1, none | 95.7% | |||
| GPT-4o mini | Therapeutic prompt v0, none | 95.6% | |||
| Therapeutic prompt v0, none | 95.6% | ||||
| Default system prompt v1, none | 95.5% | ||||
| Grok 4.3 | Default system prompt v1, high | 95.4% | |||
| Therapeutic prompt v0, default | 95.3% | ||||
| Therapeutic prompt v0, low | 95% | ||||
| Default system prompt v1, none | 95% | ||||
| Granite 4.1 8B | Default system prompt v1, none | 94.7% | |||
| Grok 4.3 | Default system prompt v1, default_low | 94.5% | |||
| Grok 4.5 | Default system prompt v1, high | 94.5% | |||
| Therapeutic prompt v0, low | 94.5% | ||||
| Therapeutic prompt v0, high | 94.1% | ||||
| Grok 4.3 | Default system prompt v1, none | 93.9% | |||
| Grok 4.3 | Therapeutic prompt v0, none | 93.7% | |||
| Therapeutic prompt v0, low | 93.7% | ||||
| Grok 4.5 | Therapeutic prompt v0, high | 93.6% |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official K-Bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-5.5 | Default system prompt v1, low | 99% | |||
| GPT-5.2 | Therapeutic prompt v0, none | 99% | |||
| GPT-5.5 | Default system prompt v1, default_medium | 98.9% | |||
| GPT-5.5 | Therapeutic prompt v0, low | 98.7% | |||
| GPT-5.5 | Therapeutic prompt v0, default_medium | 98.7% | |||
| GPT-5.2 | Therapeutic prompt v0, high | 98.6% | |||
| Kimi K2.5 | Default system prompt v1, default | 98.6% | |||
| GPT-5.2 | Default system prompt v1, none | 98.6% | |||
| Claude Opus 4.8 | Default system prompt v1, low | 98.6% | |||
| Claude Fable 5 | Default system prompt v1, high | 98.6% | |||
| GPT-5.2 | Default system prompt v1, high | 98.5% | |||
| GPT-5.4 | Therapeutic prompt v0, low | 98.5% | |||
| Claude Fable 5 | Default system prompt v1, low | 98.5% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, none | 98.5% | |||
| Kimi K2.5 | Therapeutic prompt v0, default | 98.5% | |||
| Kimi K2.6 | Default system prompt v1, default | 98.4% | |||
| GPT-5.4 | Default system prompt v1, none | 98.4% | |||
| GPT-5.5 | Therapeutic prompt v0, none | 98.4% | |||
| GPT-5.4 | Default system prompt v1, low | 98.4% | |||
| GPT-5.4 | Therapeutic prompt v0, high | 98.4% | |||
| GPT-5.4 | Therapeutic prompt v0, none | 98.4% | |||
| GPT-5.5 | Therapeutic prompt v0, high | 98.4% | |||
| Claude Haiku 4.5 | Default system prompt v1, high | 98.3% | |||
| Claude Opus 4.8 | Default system prompt v1, high | 98.3% | |||
| Grok 4.20 | Default system prompt v1, none | 98.3% | |||
| GPT-5.5 | Default system prompt v1, high | 98.3% | |||
| Claude Haiku 4.5 | Default system prompt v1, low | 98.2% | |||
| Kimi K2.6 | Default system prompt v1, high | 98.2% | |||
| GPT-5.4 | Default system prompt v1, high | 98.2% | |||
| Default system prompt v1, default_minimal | 98.1% | ||||
| Kimi K2.6 | Therapeutic prompt v0, default | 98.1% | |||
| Claude Opus 4.8 | Default system prompt v1, none | 98.1% | |||
| Claude Fable 5 | Therapeutic prompt v0, low | 98.1% | |||
| GPT-5.5 | Default system prompt v1, none | 98% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, none | 98% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, low | 98% | |||
| GPT-5.6 Sol | Default system prompt v1, none | 98% | |||
| GPT-5.6 Terra | Default system prompt v1, default_medium | 98% | |||
| GPT-5.6 Terra | Default system prompt v1, none | 97.9% | |||
| Claude Opus 4.8 | Therapeutic prompt v0, low | 97.9% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, default_medium | 97.8% | |||
| Claude Opus 4.8 | Therapeutic prompt v0, none | 97.8% | |||
| Claude Haiku 4.5 | Therapeutic prompt v0, high | 97.8% | |||
| Claude Fable 5 | Therapeutic prompt v0, high | 97.8% | |||
| Therapeutic prompt v0, default_medium | 97.7% | ||||
| Claude Haiku 4.5 | Default system prompt v1, none | 97.7% | |||
| GPT-5.6 Terra | Therapeutic prompt v0, high | 97.7% | |||
| Therapeutic prompt v0, high | 97.7% | ||||
| Default system prompt v1, none | 97.6% | ||||
| GPT-5.6 Sol | Therapeutic prompt v0, none | 97.6% | |||
| Therapeutic prompt v0, high | 97.6% | ||||
| GPT-5.6 Sol | Default system prompt v1, default_medium | 97.6% | |||
| GPT-5.6 Luna | Default system prompt v1, none | 97.6% | |||
| Therapeutic prompt v0, high | 97.5% | ||||
| Claude Opus 4.8 | Therapeutic prompt v0, high | 97.5% | |||
| GPT-5.6 Sol | Default system prompt v1, high | 97.5% | |||
| GLM-5.2 | Therapeutic prompt v0, high | 97.5% | |||
| Default system prompt v1, high | 97.5% | ||||
| Therapeutic prompt v0, default | 97.5% | ||||
| Therapeutic prompt v0, default_medium | 97.5% | ||||
| GPT-5.6 Terra | Default system prompt v1, high | 97.5% | |||
| GPT-5.6 Luna | Therapeutic prompt v0, none | 97.4% | |||
| GLM-5.2 | Default system prompt v1, default_high | 97.4% | |||
| Default system prompt v1, low | 97.4% | ||||
| Default system prompt v1, high | 97.3% | ||||
| Grok 4.20 | Therapeutic prompt v0, none | 97.3% | |||
| Kimi K2.6 | Therapeutic prompt v0, high | 97.3% | |||
| Grok 4.20 | Therapeutic prompt v0, high | 97.2% | |||
| Default system prompt v1, default_medium | 97.2% | ||||
| Default system prompt v1, none | 97.2% | ||||
| Therapeutic prompt v0, default_minimal | 97.2% | ||||
| Default system prompt v1, high | 97.2% | ||||
| Therapeutic prompt v0, none | 97.1% | ||||
| GPT-5.6 Luna | Default system prompt v1, default_medium | 97.1% | |||
| Default system prompt v1, default | 97% | ||||
| Therapeutic prompt v0, none | 97% | ||||
| GPT-5.6 Luna | Default system prompt v1, high | 97% | |||
| GPT-5.6 Sol | Therapeutic prompt v0, default_medium | 97% | |||
| Therapeutic prompt v0, none | 97% | ||||
| GPT-5.6 Sol | Therapeutic prompt v0, high | 97% | |||
| Default system prompt v1, default_medium | 97% | ||||
| Default system prompt v1, none | 97% | ||||
| Grok 4.20 | Default system prompt v1, high | 96.8% | |||
| Default system prompt v1, low | 96.8% | ||||
| Therapeutic prompt v0, none | 96.8% | ||||
| GPT-5.6 Luna | Therapeutic prompt v0, default_medium | 96.7% | |||
| GPT-5.6 Luna | Therapeutic prompt v0, high | 96.6% | |||
| Grok 4.3 | Therapeutic prompt v0, high | 96.5% | |||
| Therapeutic prompt v0, none | 96.4% | ||||
| Default system prompt v1, high | 96.3% | ||||
| Granite 4.1 8B | Therapeutic prompt v0, none | 96% | |||
| Grok 4.3 | Therapeutic prompt v0, default_low | 96% | |||
| Default system prompt v1, none | 95.9% | ||||
| Default system prompt v1, default | 95.7% | ||||
| GPT-4o mini | Default system prompt v1, none | 95.7% | |||
| GPT-4o mini | Therapeutic prompt v0, none | 95.6% | |||
| Therapeutic prompt v0, none | 95.6% | ||||
| Default system prompt v1, none | 95.5% | ||||
| Grok 4.3 | Default system prompt v1, high | 95.4% | |||
| Therapeutic prompt v0, default | 95.3% | ||||
| Therapeutic prompt v0, low | 95% | ||||
| Default system prompt v1, none | 95% | ||||
| Granite 4.1 8B | Default system prompt v1, none | 94.7% | |||
| Grok 4.3 | Default system prompt v1, default_low | 94.5% | |||
| Grok 4.5 | Default system prompt v1, high | 94.5% | |||
| Therapeutic prompt v0, low | 94.5% | ||||
| Therapeutic prompt v0, high | 94.1% | ||||
| Grok 4.3 | Default system prompt v1, none | 93.9% | |||
| Grok 4.3 | Therapeutic prompt v0, none | 93.7% | |||
| Therapeutic prompt v0, low | 93.7% | ||||
| Grok 4.5 | Therapeutic prompt v0, high | 93.6% |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official K-Bench leaderboard
GPT-4o mini on the official leaderboard passed 85% in July 2024
Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official K-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6%
- Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts.
Gemini 4 Carbon 0%
- Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior.
Inkling 2 Mini 8.30%
- Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions.
Inkling 2 Small 4.70%
- Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior.
Inkling 2 Mini 9.80%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (33 entries).
- Pass Models scored 28 models we track, plus 5 other entries on its leaderboard.
- Pass Newest score The newest is from July 2026, for GPT-5.6 Luna.
- Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Claude Fable 5 (#12).
- Note Headroom Saturated. The best is 99%, past 85% since July 2024.
Leaderboard import 3 pass
- Pass Importer We read the board on k-bench.ai: 111 rows for models we track and 5 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.